Bolt Forge is an open-model agent within Bolt.new that enables browser-based application building on open-source AI models with automated session de-identification.
What the product does and how it is positioned
Bolt Forge is an agent integrated into the Bolt.new development platform that runs exclusively on open-source language models. Sitting alongside the Standard and Max agents in the agent picker, it provides an environment for testing ideas, drafting code, and building applications using models like GLM 5.3 Flash, GLM 5.3, Kimi K3, and DeepSeek v4 Pro.
The agent executes projects directly in the browser through WebContainers and reserved infrastructure. Builders can opt in to share de-identified development sessions, including prompts, tool calls, and fix traces, which Bolt licenses to AI research partners like Arcee AI to train subsequent open-source models.
Source-supported ways to use the product
Draft initial application concepts, experiment with new architectures, and iterate through design variations using open models.
Write code, assemble project structures, and test live web applications locally inside the WebContainers browser runtime.
The documented workflow, where available
Choose Forge from the agent picker interface alongside the Standard and Max agents.
Review the data-sharing terms and complete the one-tap opt-in prompt to enable the open-model environment.
Develop and modify the project using open models, drawing from the monthly allocation bar without daily rate caps.
The platform automatically strips sensitive information, secrets, and personal identifiers before export.
Sanitized build traces are compiled into datasets licensed to external AI developers to train new open models.
Bolt Forge combines reserved server hardware with in-browser execution powered by WebContainers technology. By executing application code and package management locally inside the user browser, Forge avoids provisioning remote cloud servers for every individual build cycle, supporting low-latency development loops.
The agent incorporates a selectable roster of open-source models. Standard options include GLM 5.3 Flash and GLM 5.3, with experimental support for Kimi K3 and DeepSeek v4 Pro. When tested on the Bolt Build Index, which benchmarks real application completion, Forge open models reached 91 percent of the top proprietary model score.
Checks to run with your own material and workflow
What was checked and when
Answers based on the source-checked product record
No. Data sharing in Forge is strictly opt-in via a consent screen, and declining keeps the workspace in Standard or Max with no session sharing.
When the allocation bar reaches 100 percent, the platform automatically switches the session back to the Standard agent rather than charging overages.
No. Sessions originating from the EEA, the UK, and Switzerland are excluded from dataset licensing and AI model training.
Forge models are classified as experimental inside Bolt, and official guidance suggests duplicating projects and keeping complex production workloads in Standard or Max.
Users can halt future session collection by switching back to Standard or Max, and can request removal of prior data by emailing [email protected].
